consolidation fluxnet hybrid

[lazarusA]

Things to consider:
I'm using

• for fAPAR with the cVegLeafBareFrac approach where with consider the bare soil See:

  SINDBAD/examples/exp_fluxnet_hybrid/settings_fluxnet_hybrid/model_structure.json

  Line 100 in b701b3f

  "approach": "cVegLeafBareFrac"

  • which outputs fAPAR_bare used later in the cost function.
• cBiomass
  • note also the use of

    SINDBAD/examples/exp_fluxnet_hybrid/settings_fluxnet_hybrid/model_structure.json

    Line 85 in b701b3f

    "approach": "treeGrass"

[dr-ko]

    "hybrid.ml_training.options.batch_size" => 16,
    "hybrid.ml_training.which_fold" => 10000,
    "hybrid.ml_training.fold_path" => "blablabla",
    "hybrid.ml_model.options.n_layers" => 300,
    "hybrid.ml_model.options.n_neurons" => 5000,

[dr-ko]

This works as we need.

[lazarusA]

good, they work.

[lazarusA]

we also need the syntax for the experiment output, currently is relative to the project's directory, but in the cluster we want to send the output to an absolute path. In raven I was doing this:

SINDBAD/examples/exp_fluxnet_hybrid/exp_fluxnet_hybrid_clean_la.jl

Line 229 in b701b3f

remote_raven = "/ptmp/lalonso/HybridOutputALL/HyALL_ALL_fold_$(_nfold)_nlayers_$(nlayers)_n_neurons_$(n_neurons)_batch_size_$(batch_size)/"

and then passing that to path_experiment=checkpoint_path .

[dr-ko]

Its better to change the whole output directory to avoid overwriting in parallelized runs as,

n_fold = 5
n_layers = 3
n_neurons = 32
batch_size = 32
replace_info = Dict(
    "experiment.basics.name" => "hybrid_fold_$(n_fold)_nlayers_$(n_layers)_n_neurons_$(n_neurons)_batch_size_$(batch_size)",
)

Then the checkpoints will be saved in:
"/Users/skoirala/research/RnD/SINDBAD/examples/exp_WROASTED/../exp_fluxnet_hybrid/output_FLUXNET_hybrid_fold_5_nlayers_3_n_neurons_32_batch_size_32/hybrid/training_checkpoints"

[dr-ko]

and the base path for output can be set by a full path as:

    "experiment.model_output.path" => "/Users/youruser/yourdirectory/tmp_abspath_test",

or in the experiment.json. Note that you can also add the unique info of folds and layers to the base path, and not use the information in experiment.basics.name

[lazarusA]

yes, I would rather use an absolute path. Thanks. I will update accordingly.

[lazarusA]

@dr-ko spinup per site should be taken care of here, I think. Not sure if this is part of prepareHybrid. We need to double check this.

[dr-ko]

looks just like the forcing and not sequence.

[lazarusA]

@dr-ko spinup/costOptions per site should be taken care of here, I think.

[dr-ko]

also, cost per site.

[lazarusA]

There was a small error for Polyester, now is working. And is really fast!

> julia -t 24

julia > include("exp_fluxnet_hybrid_replace_r.jl")